A method and device for training a 3D radio map model, an electronic device, and a storage medium

Abstract

This application provides a training method, apparatus, electronic device, and storage medium for a 3D radio map model. The method includes inputting the noise latent variables, multimodal environmental data, and diffusion time steps of the target radio map of the sample area into a denoising backbone network, performing a denoising iterative process to output a predicted noise residual vector. A low-rank adapter is injected into the key projection layer of the denoising backbone network. In each iteration, the original weights of the denoising backbone network are frozen, and only the low-rank adapter is trained to determine the current training weights of the denoising backbone network. In each iteration, the predicted noise residual vector is mapped back to the physical parameter space through the frozen decoder to output a reconstructed radio map, thereby obtaining a trained radio map model. The radio map includes at least four modal data: path loss, angle of arrival, time of arrival, and delay.

Description

Technical Field

[0001] This application relates to the field of radio technology, and more specifically, to a training method, apparatus, electronic device, and storage medium for a 3D radio map model. Background Technology

[0002] With the rapid development of computer graphics and wireless communication technologies, ray tracing technology has become an important tool for radio propagation modeling.

[0003] Existing ray tracing technologies are mainly divided into two categories: physics-based deterministic algorithms and data-driven deep learning methods. Deterministic algorithms achieve millimeter-level accuracy in simple scenarios by precisely solving Maxwell's equations or approximating the wave equation. These methods require building detailed 3D environment models, and their computational complexity is proportional to the cube of the number of scene primitives. In city-level scenarios, a single prediction often requires several hours of computation, which is completely unacceptable for real-time applications such as 5G network optimization.

[0004] In recent years, deep learning technology has been introduced into the field of ray tracing to improve computational efficiency. Mainstream methods employ convolutional neural network architectures such as 3D-UNet or Pix2Pix, directly predicting field strength distribution from scene geometry through end-to-end training. While these models significantly improve computational speed, due to the local receptive field characteristics of convolutional kernels, they struggle to model the long-distance propagation effects of electromagnetic waves in large-scale scenes. This results in problems such as failing to accurately capture multiple reflections across city blocks, incorrectly handling diffraction fields at building edges, and omitting contributions from long-distance scattering. More seriously, existing data-driven models often produce predictions that violate the fundamental laws of the wave equation, such as allowing signals to penetrate metal walls without attenuation or exhibiting non-physical field strength abrupt changes in free space. Summary of the Invention

[0005] The purpose of this application is to provide a training method, apparatus, electronic device, and storage medium for a 3D radio map model, in order to solve the technical problem of inaccurate predictions in existing radio map models.

[0006] In a first aspect, the present invention provides a training method for a 3D radio map model. The method includes inputting the noise latent variables, multimodal environmental data, and diffusion time steps of the target radio map of the sample area into a denoising backbone network, performing a denoising iterative process to output a predicted noise residual vector. A low-rank adapter is injected into the key projection layer of the denoising backbone network. In each iteration, the original weights of the denoising backbone network are frozen, and only the low-rank adapter is trained to determine the current training weights of the denoising backbone network. In each iteration, the predicted noise residual vector is mapped back to the physical parameter space through the frozen decoder, and a reconstructed radio map is output to obtain a trained radio map model. The radio map includes at least four modal data: path loss, angle of arrival, time of arrival, and delay.

[0007] In an optional implementation, the noise latent variables of the target radio map for the sample area are obtained by: The target radio map is compressed using an asymmetric 3D variational autoencoder to obtain a compressed coding vector; We want to add Gaussian noise to the compressed encoding vector to obtain the noise latent variable corresponding to the time step.

[0008] In an optional implementation, the multimodal environmental data includes at least building semantic segmentation maps, building height maps, and transmitter locations. The step of inputting the noise latent variables of the target radio map of the sample area, the multimodal environmental data, and the diffusion time step into the denoising backbone network specifically includes: A lightweight 3D convolutional encoder is used to map the building semantic segmentation map, building height map, and transmitter position into an environmental feature map with the same spatial resolution as the noise latent variable. The noise latent variables and environmental feature maps are concatenated to obtain a fused input, which is then input into the denoising backbone network.

[0009] In an optional implementation, in each iteration, the predicted noise residual vector is output as follows: Based on the fused input, a token sequence is formed; The token sequence is sequentially input into multiple cascaded diffusion-based 3D generation modules to obtain the predicted noise residual vector.

[0010] In an optional implementation, the 3D generation module based on the diffusion Transformer includes a QKV projection layer, a multi-head attention layer, an output projection layer, and a feedforward output layer, with a low-rank adapter connected to the QKV projection layer. In each iteration, the token sequence is input into the low-rank adapter through the QKV projection layer, and the currently frozen weight matrix of the QKV projection layer is input into the low-rank adapter to obtain the intermediate noise residual vector output by the low-rank adapter. Based on the intermediate noise residual vector output by the low-rank adapter, the physical loss corresponding to this iteration is determined. Based on the physical loss and mean squared error loss corresponding to this iteration, the parameters of the low-rank adapter in the next iteration are adjusted as the total loss.

[0011] In an optional implementation, the physical loss is determined in each iteration by the following method: : ; ; ; in, For spatial smoothing terms, For the range of values, , For the corresponding weighting coefficients, For discrete 3D Laplacian operators, For air mask, This is the radio map output for this round. For Hadamard product operations, This is a linear activation operation.

[0012] In an optional implementation, the step of mapping the predicted noise residual vector back to the physical parameter space using the frozen decoder and outputting the reconstructed radio map in each iteration specifically includes: Estimated diffusion time Clean latent variables at time : ; in, for The prediction noise residual vector at time step [time]. This represents the noise control factor for the diffusion process.

[0013] In an optional implementation, it further includes: The multimodal environmental data and diffusion time corresponding to the area to be generated are input into the trained radio map model to obtain the radio map of the area to be generated output by the radio map model.

[0014] Secondly, the present invention provides a training device for a 3D radio map model. The device includes a training module for inputting noise latent variables, multimodal environmental data, and diffusion time steps of a target radio map of a sample area into a denoising backbone network, performing a denoising iterative process to output a predicted noise residual vector. A low-rank adapter is injected into the key projection layer of the denoising backbone network. In each iteration, the original weights of the denoising backbone network are frozen, and only the low-rank adapter is trained to determine the current training weights of the denoising backbone network. In each iteration, the predicted noise residual vector is mapped back to the physical parameter space using the frozen decoder, outputting a reconstructed radio map to obtain the trained radio map model. Radio maps include at least four modal data: path loss, angle of arrival, time of arrival, and delay.

[0015] Thirdly, the present invention provides an electronic device, comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the training method for any of the 3D radio map models described in the foregoing embodiments.

[0016] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of a training method for any of the 3D radio map models described in the foregoing embodiments.

[0017] This application provides a training method, apparatus, electronic device, and storage medium for a 3D radio map model. The method includes inputting the noise latent variables, multimodal environmental data, and diffusion time steps of the target radio map of a sample area into a denoising backbone network, performing an iterative denoising process to output a predicted noise residual vector. A low-rank adapter is injected into the key projection layer of the denoising backbone network. In each iteration, the original weights of the denoising backbone network are frozen, and only the low-rank adapter is trained to determine the current training weights of the denoising backbone network. In each iteration, the predicted noise residual vector is mapped back to the physical parameter space through the frozen decoder, outputting a reconstructed radio map to obtain a trained radio map model. The radio map includes at least four modal data: path loss, angle of arrival, time of arrival, and delay. This application, by combining LoRA technology, eliminates the need for full fine-tuning of a large model; only a very small number of parameters need to be updated to achieve physical property adaptation, greatly reducing computational requirements. Through Tweedie estimation and physical loss, the model output is forced to conform to the constraints of Maxwell's equations, eliminating non-physical "signal illusions." Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A framework diagram of a training method for a 3D radio map model provided in an embodiment of this application; Figure 2 A schematic diagram of the structure of a training device for a 3D radio map model provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0021] Example 1 Figure 1 This is a framework diagram of a training method for a 3D radio map model provided in an embodiment of this application. (See diagram for example.) Figure 1 As shown, in one embodiment of this application, a generative framework based on PhysDiT (Physics-Aware DiffusionTransformer) is provided, which aims to solve the problem that existing data-driven models lack physical consistency and long-distance dependency modeling capabilities.

[0022] The model was trained and tested using the multimodal, 3D urban dataset UrbanRadio3D. The following example illustrates the generation of a 3D radio map in an urban environment (including multimodal data such as path loss, angle of arrival (DoA), and time of arrival (ToA)).

[0023] To preserve high-frequency geometric details while reducing computational complexity, one embodiment of this application does not directly manipulate the original pixel space, but instead employs a "compression-fusion" strategy. In a feasible embodiment, the noise latent variables of the target radio map of the sample area can be obtained in the following way: The target radio map is compressed using an asymmetric 3D variational autoencoder to obtain a compressed coding vector. Gaussian noise is then added to the compressed coding vector to obtain the noise latent variable corresponding to each time step.

[0024] Specifically, pre-trained asymmetric 3D variational autoencoders (3D-VAEs) can be used. High-dimensional target radio map volume (in Corresponding to four modes, Compression of spatial dimensions into a compact latent representation .

[0025] During the training phase, the potential representations are... Add Gaussian noise Get time step noise latent variables .

[0026] The noise latent variables of the target radio map of the sample area, multimodal environmental data, and diffusion time steps are input into the denoising backbone network to perform an iterative denoising process, outputting a predicted noise residual vector. The multimodal environmental data includes at least building semantic segmentation maps, building height maps, and transmitter locations.

[0027] The steps of inputting the noise latent variables, multimodal environmental data, and diffusion time steps of the target radio map of the sample area into the denoising backbone network specifically include: A lightweight 3D convolutional encoder maps the building semantic segmentation map, building height map, and transmitter location to an environmental feature map with the same spatial resolution as the noise latent variables. The noise latent variables and environmental feature maps are then concatenated to obtain a fused input, which is fed into the denoising backbone network.

[0028] Specifically, environmental conditions can be taken into account. Includes: 3D building semantic segmentation maps Building height map and transmitter location .

[0029] Construct a lightweight 3D convolutional encoder to map the above environmental conditions to noise latent variables. With the same spatial resolution, environmental feature maps are obtained. .

[0030] Unlike the standard DiT model which injects conditions through cross-attention, one embodiment of this application employs an early fusion strategy. Noise latent variables are directly introduced into the backbone network before being fed into the backbone network. Environmental feature map The input is obtained by concatenating the data along the channel dimension. This design forces the model to strictly align geometric priors with the generated signal field at the pixel level, avoiding spatial information "drift" that may be caused by cross-attention.

[0031] In each iteration, the predicted noise residual vector can be output in the following way: Based on the fused input, a token sequence is formed. The fused input can be... Divide into non-overlapping 3DPatches (e.g., size 10 ... Each patch is flattened and mapped to a one-dimensional token sequence through a linear projection layer, then fed into the subsequent network after incorporating learnable positional encodings.

[0032] Next, the token sequence is sequentially input into multiple cascaded diffusion-based 3D generation modules to obtain the prediction noise residual vector.

[0033] The 3D generation module based on diffusion Transformer includes a QKV projection layer, a multi-head attention layer, an output projection layer, and a feedforward output layer.

[0034] Understandably, in one feasible implementation, a 3D Diffusion Transformer (DiT) is used as the denoising backbone network. It is used to capture the global propagation characteristics of electromagnetic waves in complex urban environments (such as multipath reflection over long distances).

[0035] In a specific embodiment, the denoising backbone network can be composed of... It consists of stacked DiT-3D Blocks (3D generation modules based on diffusion Transformer).

[0036] Each block receives the embedding vector at time step t, first using MLP to regress the scaling factor γ and translation factor β, and then normalizing the input token. This module relies only on time step information and does not introduce spatial conditions, ensuring that the spatial structure information is dominated by the concatenated input from the previous step.

[0037] Subsequently, the Multi-Head Self-Attention (MHSA) layer calculates the global attention weights between tokens to capture the geometric dependencies across the entire scene.

[0038] The feedforward output layer is used to perform pointwise feature transformation.

[0039] Furthermore, residual connections are set at the tail of both the multi-head attention layer and the feedforward output layer.

[0040] In other embodiments of this application, the DiT in the denoising backbone network can also be replaced with Video VisionTransformer (ViViT) or other 3D Attention architectures.

[0041] Furthermore, in order to inject physical constraints without disrupting the priors generated by pre-training, a low-rank adapter can be connected to the QKV projection layer as a physical corrector.

[0042] In each iteration, the token sequence is input into the low-rank adapter via a QKV projection layer, and the currently frozen weight matrix of the QKV projection layer is also input into the low-rank adapter to obtain the intermediate noise residual vector output by the low-rank adapter. Based on the intermediate noise residual vector output by the low-rank adapter, the physical loss corresponding to this iteration is determined. The physical loss and mean squared error loss corresponding to this iteration are used as the total loss to adjust the parameters of the low-rank adapter for the next iteration.

[0043] The key projection layers of the denoising backbone network are injected with low-rank adapters. These low-rank adapters can also be replaced with Adapter, Prefix Tuning, or other Parameter Efficient Fine-Tuning (PEFT) modules.

[0044] In each iteration, the original weights of the denoising backbone network are frozen, and only the low-rank adapter is trained to determine the current training weights of the denoising backbone network.

[0045] In each iteration, the predicted noise residual vector is mapped back to the physical parameter space through the frozen decoder, and the reconstructed radio map is output to obtain the trained radio map model. The radio map includes at least four modal data: path loss, angle of arrival, time of arrival, and delay.

[0046] Specifically, during the physical adaptation training phase, all original weight parameters of the pre-trained DiT backbone network are completely frozen. (Including weights in MHSA and FFN).

[0047] Only in the key linear projection layer of the Transformer layer (especially the Query, Key, Value projection matrix) and output matrix Parallel injection of trainable low-rank branches.

[0048] The forward propagation formula for LoRA is: ; in, For frozen pre-trained weights, This is a dimension-reduced matrix (initialized using Gaussian). This is an upgraded matrix (initialized with all zeros). Rank (e.g.) or ), This is the scaling factor.

[0049] Thus, compared with the full parameter training in the prior art, this application only requires a small number of parameters to achieve the adaptation of physical characteristics.

[0050] Furthermore, since the diffusion model processes noisy data at intermediate time steps... Directly applying physical constraints is ineffective. This invention proposes to calculate physical losses by estimating "clean" physical quantities using the Tweedie formula.

[0051] The noise predicted by the current model can be utilized. Estimate based on Tweedie formula Clean latent variables at time : ; in These are the noise scheduling parameters for the diffusion process.

[0052] Using the frozen VAE decoder The estimated latent variables Mapping back to the physical parameter space yields the reconstructed radio map. .

[0053] Physical loss function (physical loss) The construction steps may include: In physical space The above calculates the explicit physical constraint loss, including (1) Spatial smoothing term ( ): Using the discrete 3D Laplace operator ( ) Calculate the free space region (by air mask) The second derivative of the index (indicating the second derivative) constrains the smoothness and rotational invariance of electromagnetic wave propagation: ; (2) Numerical range item ( ): Punishment exceeding the range of physically meaningful values ​​(e.g.) The predicted value (normalized interval) eliminates non-physical "illusions" such as enhanced signal penetration through walls: ; in, For spatial smoothing terms, For the range of values, , For the corresponding weighting coefficients, For discrete 3D Laplacian operators, For air mask, This is the radio map output for this round. For Hadamard product operations, This is a linear activation operation.

[0054] The total physical loss can be expressed as: .

[0055] Therefore, the overall training objective is: . This represents the mean squared error loss.

[0056] During reverse propagation, blocking DiT backbone parameters Gradient update, forcing the gradient of physical loss Only flow to LoRA parameter ( ).

[0057] This mechanism allows the LoRA module to be specifically responsible for learning "how to correct physical errors," while the backbone network maintains its general data distribution generation capabilities.

[0058] This application provides a training method for a 3D radio map model. Compared to CNN methods such as RadioUNet, this method utilizes the global attention mechanism of DiT to accurately simulate long-distance multipath effects in urban canyons. By using Tweedie estimation and physics loss, the model output is forced to conform to the constraints of Maxwell's equations, and non-physical "signal illusions" can be eliminated. Combined with LoRA technology, it eliminates the need for full fine-tuning of a large model; only a very small number of parameters need to be updated to adapt to physical properties, greatly reducing computational requirements.

[0059] Example 2 In one embodiment of this application, multimodal environmental data and diffusion time corresponding to the area to be generated are input into a trained radio map model to obtain a radio map of the area to be generated output by the radio map model. Specifically, the area to be generated can be a city.

[0060] Example 3 Figure 2 This is a schematic diagram of the structure of a training device for a 3D radio map model provided in an embodiment of this application. Figure 2 As shown, based on the same inventive concept, this application also provides a training device for a 3D radio map model. The device includes a training module 200, which is used to input the noise latent variables, multimodal environmental data, and diffusion time steps of the target radio map of the sample area into a denoising backbone network, perform a denoising iterative process, and output a predicted noise residual vector. A low-rank adapter is injected into the key projection layer of the denoising backbone network. In each iteration, the original weights of the denoising backbone network are frozen, and only the low-rank adapter is trained to determine the current training weights of the denoising backbone network. In each iteration, the predicted noise residual vector is mapped back to the physical parameter space using the frozen decoder, outputting a reconstructed radio map to obtain the trained radio map model. Radio maps include at least four modal data: path loss, angle of arrival, time of arrival, and delay.

[0061] In a preferred embodiment, the noise latent variables of the target radio map of the sample area are obtained by the following method: The target radio map is compressed using an asymmetric 3D variational autoencoder to obtain a compressed coding vector; We want to add Gaussian noise to the compressed encoding vector to obtain the noise latent variable corresponding to the time step.

[0062] In a preferred embodiment, the multimodal environmental data includes at least a building semantic segmentation map, a building height map, and transmitter locations. The step of inputting the noise latent variables of the target radio map of the sample area, the multimodal environmental data, and the diffusion time step into the denoising backbone network specifically includes: A lightweight 3D convolutional encoder is used to map the building semantic segmentation map, building height map, and transmitter position into an environmental feature map with the same spatial resolution as the noise latent variable. The noise latent variables and environmental feature maps are concatenated to obtain a fused input, which is then input into the denoising backbone network.

[0063] In a preferred embodiment, in each iteration, the predicted noise residual vector is output in the following manner: Based on the fused input, a token sequence is formed; The token sequence is sequentially input into multiple cascaded diffusion-based 3D generation modules to obtain the predicted noise residual vector.

[0064] In a preferred embodiment, the 3D generation module based on diffusion Transformer includes a QKV projection layer, a multi-head attention layer, an output projection layer, and a feedforward output layer. A low-rank adapter is connected to the QKV projection layer. In each iteration, the token sequence is input into the low-rank adapter through the QKV projection layer, and the currently frozen weight matrix of the QKV projection layer is input into the low-rank adapter to obtain the intermediate noise residual vector output by the low-rank adapter. Based on the intermediate noise residual vector output by the low-rank adapter, the physical loss corresponding to this iteration is determined. Based on the physical loss and mean squared error loss corresponding to this iteration, the parameters of the low-rank adapter in the next iteration are adjusted as the total loss.

[0065] In a preferred embodiment, the physical loss is determined in each iteration by the following method: : ; ; ; in, For spatial smoothing terms, For the range of values, , For the corresponding weighting coefficients, For discrete 3D Laplacian operators, For air mask, This is the radio map output for this round. For Hadamard product operations, This is a linear activation operation.

[0066] In a preferred embodiment, the step of mapping the predicted noise residual vector back to the physical parameter space using the frozen decoder and outputting the reconstructed radio map in each iteration specifically includes: Estimated diffusion time Clean latent variables at time : ; in, for The prediction noise residual vector at time step [time]. This represents the noise control factor for the diffusion process.

[0067] In a preferred embodiment, a prediction module is also provided, which is used to input the multimodal environmental data and diffusion time corresponding to the area to be generated into the trained radio map model to obtain the radio map of the area to be generated output by the radio map model.

[0068] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 3 As shown, the electronic device 300 includes a processor 310, a memory 320, and a bus 330.

[0069] The memory 320 stores machine-readable instructions executable by the processor 310. When the electronic device 300 is running, the processor 310 and the memory 320 communicate via the bus 330. When the machine-readable instructions are executed by the processor 310, they can perform the operations described above. Figure 1 The steps of a training method for a 3D radio map model shown in the method embodiment are described in detail in the method embodiment, and will not be repeated here.

[0070] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The steps of a training method for a 3D radio map model shown in the method embodiment are described in detail in the method embodiment, and will not be repeated here.

[0071] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0072] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0073] Furthermore, the units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0074] Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0075] It should be noted that if the function is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0076] In this document, relational terms such as first and second are used only to distinguish one entity or operation from another entity or operation, without necessarily requiring or implying any such actual relationship or order between these entities or operations.

[0077] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A training method for a 3D radio map model, characterized in that, The method includes: The noise latent variables of the target radio map of the sample area, multimodal environmental data, and diffusion time steps are input into the denoising backbone network to perform a denoising iterative process, outputting a predicted noise residual vector. A low-rank adapter is injected into the key projection layer of the denoising backbone network. In each iteration, the original weights of the denoising backbone network are frozen, and only the low-rank adapter is trained to determine the current training weights of the denoising backbone network. In each iteration, the predicted noise residual vector is mapped back to the physical parameter space using the frozen decoder, outputting a reconstructed radio map to obtain the trained radio map model. Radio maps include at least four modal data: path loss, angle of arrival, time of arrival, and delay.

2. The method according to claim 1, characterized in that, The noise latent variables of the target radio map for the sample area are obtained using the following method: The target radio map is compressed using an asymmetric 3D variational autoencoder to obtain a compressed coding vector; Gaussian noise is added to the compressed coding vector to obtain the noise latent variable corresponding to the time step.

3. The method according to claim 2, characterized in that, The multimodal environmental data includes at least building semantic segmentation maps, building height maps, and transmitter locations. The step of inputting the noise latent variables of the target radio map of the sample area, the multimodal environmental data, and the diffusion time step into the denoising backbone network specifically includes: A lightweight 3D convolutional encoder is used to map the building semantic segmentation map, building height map, and transmitter position into an environmental feature map with the same spatial resolution as the noise latent variable. The noise latent variable and the environmental feature map are concatenated to obtain a fused input, which is then input into the denoising backbone network.

4. The method according to claim 3, characterized in that, In each iteration, the predicted noise residual vector is output as follows: Based on the fused input, a token sequence is formed; The token sequence is sequentially input into multiple cascaded diffusion-based 3D generation modules to obtain the predicted noise residual vector.

5. The method according to claim 4, characterized in that, The 3D generation module based on diffusion Transformer includes a QKV projection layer, a multi-head attention layer, an output projection layer, and a feedforward output layer. A low-rank adapter is connected to the QKV projection layer. In each iteration, the token sequence is input into the low-rank adapter through the QKV projection layer, and the currently frozen weight matrix of the QKV projection layer is input into the low-rank adapter to obtain the intermediate noise residual vector output by the low-rank adapter. Based on the intermediate noise residual vector output by the low-rank adapter, the physical loss corresponding to this iteration is determined. Based on the physical loss and mean squared error loss corresponding to this iteration, the parameters of the low-rank adapter in the next iteration are adjusted as the total loss.

6. The method according to claim 4, characterized in that, In each iteration, the physical loss is determined as follows: : ； ； ； in, For spatial smoothing terms, For the range of values, , For the corresponding weighting coefficients, For discrete 3D Laplacian operators, For air mask, This is the radio map output for this round. For Hadamard product operations, This is a linear activation operation.

7. The method according to claim 1, characterized in that, In each iteration, the step of mapping the predicted noise residual vector back to the physical parameter space using the frozen decoder and outputting the reconstructed radio map specifically includes: Estimated diffusion time Clean latent variables at time : ； in, for The prediction noise residual vector at time step 1. This represents the noise control factor for the diffusion process.

8. The method according to claim 1, characterized in that, Also includes: The multimodal environmental data and diffusion time corresponding to the area to be generated are input into the trained radio map model to obtain the radio map of the area to be generated output by the radio map model.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the training method for the 3D radio map model as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the training method for the 3D radio map model as described in any one of claims 1 to 7.

Images